=Paper=
{{Paper
|id=Vol-2890/paper3
|storemode=property
|title=An Automatic Framework to Continuously Monitor Multi-Platform Information Spread
|pdfUrl=https://ceur-ws.org/Vol-2890/paper3.pdf
|volume=Vol-2890
|authors=Zhouhan Chen,Kevin Aslett,Jen Rosiere Reynolds,Juliana Freire,Jonathan Nagler,Joshua A. Tucker,Richard Bonneau
|dblpUrl=https://dblp.org/rec/conf/www/ChenARFNTB21
}}
==An Automatic Framework to Continuously Monitor Multi-Platform Information Spread==
https://ceur-ws.org/Vol-2890/paper3.pdf
An Automatic Framework to Continuously
Monitor Multi-Platform Information Spread
Zhouhan Chen, Kevin Aslett, Jen Rosiere Reynolds, Juliana Freire, Jonathan
Nagler, Joshua A. Tucker, and Richard Bonneau
New York University, New York NY 10003, USA
{zc1245,kma412,jr5505,juliana.freire,jonathan.nagler,joshua.tucker,bonneau}@nyu.edu
Abstract. Identifying and tracking the proliferation of misinformation,
or fake news, poses unique challenges to academic researchers and online
social networking platforms. Fake news increasingly traverses multiple
platforms, posted on one platform and then re-shared on another, mak-
ing it difficult to manually track the spread of individual messages. Also,
the prevalence of fake news cannot be measured by a single indicator,
but requires an ensemble of metrics that quantify information spread
along multiple dimensions. To address these issues, we propose a frame-
work called Information Tracer, that can (1) track the spread of news
URLs over multiple platforms, (2) generate customizable metrics, and
(3) enable investigators to compare, calibrate, and identify possible fake
news stories. We implement a system that tracks URLs over Twitter,
Facebook and Reddit and operationalize three impact indicators – Total
Interaction, Breakout Scale and Coefficient of Traffic Manipulation – to
quantify news spread patterns. Using a collection of human-verified false
URLs, we show that URLs from different origins have different propensi-
ties to spread to multiple platforms, cover different topics, while exhibit
similar retweet patterns. We also demonstrate how our system can dis-
cover URLs whose spread pattern deviate from the norm, and be used to
coordinate human fact-checking of news domains. Our framework pro-
vides a readily usable solution for researchers to trace information across
multiple platforms, to experiment with new indicators, and to discover
low-quality news URLs in near real-time.
Keywords: misinformation, cross platform, fake news, human computer
interaction, information flow, anomaly detection
1 Introduction
The COVID-19 pandemic has increased the consumption of news via social me-
dia. For example, [1] a recent global survey found that, since the beginning of
Copyright c 2021 for this paper by its authors. Use permitted under Creative Com-
mons License Attribution 4.0 International (CC BY 4.0). Presented at the MISINFO
2021 Workshop, held in conjunction with the 30th ACM The Web Conference, 2021,
in Ljubljana, Slovenia.
COVID-19, 43% of consumers increased time spent on YouTube, 40% on Face-
book and 23% on Twitter. As people spend more time consuming news from
online platforms, the volume of online misinformation has also increased, result-
ing in the World Health Organization declaring an Infodemic [23]. To mitigate
misinformation and promote high-quality content, it is important for us to first
understand where information originates and how it spreads. Two major tech-
nical challenges remain. First, information is often posted on one platform and
shared on another, but recent work in cross-platform news spread focus on sin-
gle events, which are ad-hoc and not scalable [21,7]. Second, there is no unified
approach to measure and quantify information spread. Different measurements
result in different estimations of misinformation prevalence. For example, [22]
points out that depending on the chosen datasets and metrics, the amount of
misinformation on Twitter can range between 1% to 70%. Measuring the preva-
lence of fake news with a single indicator is inadequate.
In this paper, we propose a framework called Information Tracer that con-
tributes three major improvements to previous work. First, we define a unified
data collection pipeline to trace and visualize data from multiple platforms.
Second, we support a multi-pronged approach that uses multiple indicators to
measure information spread. Third, we provide a user interface to enable re-
searchers to comparatively identify URLs with unusual metrics, and to facil-
itate fact-checkers by contextualizing URL spread across multiple platforms.
We implement Information Tracer to track URLs over three platforms – Twit-
ter, Facebook, and Reddit, the most popular mobile social networking plat-
forms in the United States as of September 2019 [19]. To quantify information
spread, we operationalize three impact indicators – Total Interaction, Breakout
Scale, and Coefficient of Traffic Manipulation. Finally, we create a web interface
(https://informationtracer.com/) to visualize both raw data and aggregated
statistics.
We also present three real-world applications to demonstrate the capabil-
ity of Information Tracer. In Application One, we investigate three main ques-
tions using a collection of fake news URLs from four origins (Twitter, Facebook,
YouTube, News outlets):
1. Do URLs from different origins have different likelihood to spread across
multiple platforms?
2. Do they have different Twitter retweet traffic patterns?
3. Do they cover different topics?
We find that URLs from Facebook are less likely to spread over multiple plat-
forms; URLs from different platforms cover different false stories; and there is
no significant difference in retweet patterns.
Application Two (A2) and Three (A3) include human oversight and interac-
tion, so called human-in-the-loop capability, to our framework. In A2, we demon-
strate how Information Tracer can assist humans to identify URLs whose impact
indicators deviate from the sample average. In A3, we instruct human coders to
fact-check qualities of news domains with the help of Information Tracer. We
show that our system can potentially reduce the time it takes to discover previ-
ously unknown low-quality news sites.
The paper is organized as the following: Section 2 details each component of
Information Tracer system. Section 3 applies our framework in three three real-
world settings. Section 4 discusses the limitation of our research. We examine
related work in Section 5, and conclude the paper in Section 6.
2 Information Tracer System
On a high level, Information Tracer consists of three components – data col-
lector, data aggregator, and data visualizer; these modules collect data,
generate summary statistics, and enable visualizing data respectively. Figure 1
shows the system architecture. In this section, we detail how we implement each
component.
Although we implement our framework with a particular set of configurations
that help us answer our research questions, our proposed framework is customiz-
able and users can design their own metrics to better answer other questions.
A metric can be a simple count, or a numerical output from a machine learning
model. Our framework is also extendable – users can integrate additional sources
(social media platforms, weblogs, messaging software) into the system, without
altering the overall data pipeline.
…
Fig. 1: Information Tracer Architecture. The input to the system is a valid URL.
Upon receiving the URL, our system automatically collects posts containing
this URL from designated social networking platforms, generates aggregated
statistics, and finally presents results on a web interface.
2.1 Component One: data collector
The goal of the data collector is to parse queries submitted by end users, then
collect posts that match those queries from a list of platforms. For the scope of
this paper, we restrict the query to a valid URL, and we consider three plat-
forms – Twitter, Facebook and Reddit. We focus on the URL because it has a
well-defined structure, is indexed by all three platforms, and serves as a unique
identifier of news stories.
URL sanitization and normalization. Before we make API calls to each
platform, we sanitize and normalize the input URL to maximize the number of
matched posts on our three platforms. We sanitize a URL in following ways:
– Remove prefixes http://, https:// and www.. For example, query http://www.abc.com/xyz
will become abc.com/xyz. This ensures that we match all posts that refer to
article abc.com/xyz.
– Remove query parameters. A query parameter is a substring that follows a “?”.
They are usually appended at the end of URL for tracking purposes. We strip
query parameters to normalize the input URL with a few exceptions. For ex-
ample, a standard YouTube video looks like youtube.com/watch?v=VideoID,
in which “?” is important and cannot be removed. We maintain an allowlist
of such domains.
Twitter Collection Our Twitter search is powered by Twitter Academic Track
API1 . This API provides us with access to Twitter’s full-archive tweet corpus.
As of February 17, 2021, the API imposes a cap of 10,000,000 tweets per month.
Due to this rate limit, we have to be judicious about how we collect tweets. Our
strategy is to collect influential tweets that receive a high level of interactions
such as retweets and replies, and to avoid collecting tweets with low interaction
(tweets along the “long tail”). This intuition comes from a previous study on
Twitter user characterization, which finds that a small number of influential users
control most of conversation diffusion [6]. Because the definition of “influential”
is subjective, we introduce five tunable parameters that can be specified by
users during query submission – minimum number of retweets (min retweets),
minimum number of replies (min replies), maximum number of original tweets
(max originals), maximum number of retweets (max retweets), and maximum
number of replies (max replies). The following is our data collection protocol:
1. Given a URL=q, min retweets=x, min replies=y, we construct a special URL
– https://twitter.com/search?q=min_retweets:x%20min_replies:y%20url:
q&f=live. This URL returns us matched original tweets, with at least x
retweets and y replies. We use a Python Selenium headless browser to au-
tomatically visit this URL, scroll down the page, and extract max originals
number of tweets, or until there is no result. We have to use a headless browser
to automate this process because the two search parameters (min retweets,
min replies) are not available via the API.
2. Then for each original tweet with id=TweetID, we use full-archive search
endpoint 2 to collect retweets and replies. We set query=status/TweetID to re-
trieve all quoted tweets and retweets of quoted tweets. We set query=conversation id:TweetID
to match all replies of the original tweet. We collect max retweets and max replies
number of results.
1
https://developer.twitter.com/en/solutions/academic-research
2
https://api.twitter.com/2/tweets/search/all
Our Twitter collection module is thus customizable: by tuning each threshold
one can collect more or less tweets, and adapt to different questions and API
rate limits. For example, to collect all matched retweets and replies one can set
min retweets=0, min replies=0, max originals=∞, max replies=∞, max retweets=∞.
In practice, we strongly recommend setting thresholds to avoid burning API us-
age. These settings should be application-specific, and thus, we present use cases
in Section 3.
Facebook Collection. We use Crowdtangle to collect Facebook public posts
containing the input URL. Crowdtangle is a tool that collects and aggregates
engagement data of Facebook, Instagram and Reddit posts. It provides API to
journalists and academic researchers. We use the Search API to collect Face-
book posts containing the input URL. The API returns up to 1,000 posts. To
collect influential posts, we use the sort parameter to retrieve posts with the
highest score. The score is a metric designed by Crowdtangle to indicate if a
post “overperforms.” 3 Importantly, Crowdtangle does not index every single
Facebook page. According to Crowdtangle’s documentation4 , as of February 24,
2021, more than six million Facebook pages, groups, and verified profiles are in-
dexed. This includes “all public Facebook pages and groups with more than 100K
likes, all US-based public groups with 2k+ members, and all verified profiles,”
and therefore misses private groups and pages.
Reddit Collection. Similar to Facebook data collection, we use Crowdtangle
to collect the top 1,000 Reddit posts containing the input URL sorted by the
“overperform” score. Crowdtangle indexes more than 20,000 of the most active
sub-reddits, and adds more sub-reddits on an ongoing basis.
To summarize, due to limitations from each API endpoint, we are not able
to retrieve every post that matches a query. Specifically, private posts are un-
available, and posts from less popular groups may not be indexed yet. We argue
that the omission of those low-interaction posts are acceptable because they do
not play a significant role in spreading information. From a resource allocation
perspective, storing only popular posts (cutting off the long tail) saves storage
space, and improves data processing speed.
2.2 Component Two: data aggregator
The goal of data aggregator is to distill intelligence from heterogeneous cross-
platform data sources. It achieves this goal by calculating summary statistics
to quantify information spread. In this paper, we refer to those statistics as
impact indicators, as they indicate the relative impact of a URL on one or more
platforms. Over the years many indicators have been proposed and explored. In
this paper, we operationalize three indicators – Total Interaction, Breakout
Scale [12], and Coefficient of Traffic Manipulation (CTM) [11].
We choose those measurements because they are compatible with our dataset.
Specifically, Breakout Scale requires multi-platform data to measure information
3
https://help.crowdtangle.com/en/articles/3213537-crowdtangle-codebook
4
https://help.crowdtangle.com/en/articles/1140930-what-data-is-crowdtangle-tracking
spread, CTM requires retweet data to measure Twitter traffic pattern, and Total
Interaction requires total number of interactions of every post. All three types
of data are available in our collection. We want to point out that our framework
is indicator-agnostic. The indicators we operationalized may be more helpful on
one dataset but less on another. We now introduce each indicator in detail.
Total Interaction Interaction count is a simple yet effective measurement to
quantify the popularity of a post. This metric has proven to be useful in recent
studies to quantify fake news spread during COVID-19 [15,5,10]. For each URL,
we define its total interaction as the summation of total interactions of every
Twitter, Facebook, and Reddit post. We define the post-level total interaction
as5 :
– Twitter post. The total number of retweets, replies and likes.
– Facebook post. The total number of reactions, shares and comments.
– Reddit post. The total number of upvotes and comments.
Breakout Scale Breakout Scale is originally proposed as a comparative model
for measuring and calibrating Information Operations (IOs) based on “data that
are observable, replicable, verifiable, and available from the moment they were
posted [12].” It measures how many platforms an IO percolates to, and assigns
an IO to one of six categories, as shown in Table 1.
We find the Breakout Scale framework appealing as it allows us to quantify
how many platforms a URL is popular over. To operationalize this framework,
we use total interaction as a proxy for popularity. Formally, for each URL u, we
denote the total number of interactions it receives on platform p as interactionp .
We then set a threshold t, if interactionp > t, we consider u to be popular
on platform p. The final Breakout Scale for u is the total number of popular
platforms.
Coefficient of Traffic Manipulation (CTM) We also compute fine-grained
indicators that quantify platform-specific patterns. Because we only have page
and group level statistics for Facebook and Reddit posts, we focus on summa-
rizing Twitter traffic here, for which we have full access. CTM is a comparative
model that allows one to compare different Twitter traffic flows “against measur-
able criteria and assess which of those movements appear to have been subject
to manipulation [11].”
Originally, CTM was a weighted average of three measurements: the average
number of tweets per user (m1 ), the percentage of retweets as a proportion of
total tweets (m2 ), and the proportion of tweets generated by top fifty accounts
(m3 ). After analyzing real-world Twitter traffic containing manipulated hash-
tags, the authors concluded that m1 and m3 are more informative to identify
5
Definitions for Facebook and Reddit posts are adopted from https://help.
crowdtangle.com/en/articles/1184978-crowdtangle-glossary
manipulated traffic. In our implementation, we modify and define CTM as a
tuple of two values: average number of tweets per user, and proportion of tweets
generated by top 10% accounts. We focus on percentage instead of top fifty
accounts as, in our experiments, we find tweet threads with fewer than fifty
accounts.
We want to note here that a high CTM does not always imply traffic manipu-
lation. For example, a tweet thread with high CTM could be caused by authentic
users who are engaged in the conversation and replied many times. Similarly, a
tweet corpus with low CTM might be manipulated by a sophisticated bot cam-
paign, in which each bot only creates one tweet, thus evading this metric. In
Section 3 we show how to use our system to discover the cause of high CTM.
Table 1: Definition of Breakout Scale. Our system can automatically derive Cat-
egory 0 to 3 based on data collected from three platforms.
Category Definition Can we operationalize?
0 Popular on zero platform Yes
1 Popular on one platform Yes
2 Popular on two platforms Yes
3 Popular on three or more platforms Yes
4 Popular on multiple media (online and offline) Not yet
5 Celebrity amplification Not yet
6 Require policy change Not yet
2.3 Component Three: data visualizer
Data visualization is a key element of both validating this platform and en-
abling needed human interaction. Thus we aim to facilitate real-time exchange
of cross-platform data and intelligence. We propose and implement two main
data visualizations – a summary page and an item-wise detail page.
Summary page visualization The summary page allows investigators to com-
pare, calibrate and identify data points (in our case URLs) with unusual spread
patterns. Our summary page is available at https://informationtracer.com/
intelligence. We currently use a scatter plot to visualize all three impact in-
dicators. Investigators can identify an interesting quadrant, zoom in, and click
on individual point (which represents a URL) to navigate to the detail page.
Item-wise Detail visualization page The detail page allows investigators
to visualize individual posts from different platforms, and explore how posts
interact with each other along multiple dimensions, such as temporal, network,
and contextual. Figure 2 is a rendering of one detail page that contextualizes the
spread pattern of URL www.armyfortrump.com. Those visualizations provide
answers to questions such as when the URL is shared on each platform, who
posted it, and how users who share the URL interacted with each other via
retweet and reply.
3 Real world applications
We now introduce three real-world applications (denoted as A1, A2 and A3).
A1 uses Information Tracer to understand and compare how fake news URLs
from different origins spread over three platforms. We focus on four origins for
the fake-news we will trace: Twitter, Facebook, Youtube and News domains. A2
and A3 incorporate human-in-the-loop intelligence. For A2, we use Information
Tracer to discover URLs with unusual impact indicators. For A3, we instruct
human coders to assess qualities of news domains using our system. In the rest
of the section, we first introduce our data sources, then explain each application.
3.1 Overview of datasets
Google Fact Check Dataset (abbr. Google FN ). The Google Fact Check
Dataset is a repository of false claims, fact checked by journalists around the
world. The dataset has been adopted by many fact checkers around the world,
including those verified by International Fact Checking Network (IFCN). It also
powers fact checking features behind Google Search, Google News and Bing
Search [3].
We collect all US-based claims from Google Fact Check Dataset during 2020.
To do so, we first download all claims from the web portal [2]. We sort all claims
by fact checking organizations, and manually check the origin of top 30 or-
ganizations (which account for more than 90% of all claims). We identify six
organizations that operate in the United States – politifact.com, factcheck.
org, washingtonpost.com, usatoday.com, nytimes.com, poynter.org. Then
for each claim from each organization, we examine the API structure, and ex-
tract URLs from field entry[“itemReviewed”], which are URLs that point to the
source of fake news. If the URL is archived, we run another script to extract the
original URL from archived page. In the end, we extract 1427 unique URLs.
IFCN COVID-19 Fake News Dataset (abbr. IFCN FN ). Our second
dataset contains 8,627 false claims compiled by fact checkers among IFCN. The
earliest entry is from 1/5/2020, and the latest entry is from 8/26/2020. Each
entry contains a URL that points to the source of false claim. However, there
are several special cases:
1. Shortened URLs. We write a script to resolve their final landing URLs.
2. Non-URL texts. Some URLs are plain texts such as ”web page removed”,
”There is no link,” ”It is an e-mail”. We remove those entries.
3. Duplicated URLs. We only keep the first entry.
Fig. 2: Information Tracer User Interface. This detail page visualizes the spread
pattern of URL armyfortrump.com. To understand when people share the web-
site, one can examine the scatter plot (first row). To understand who shares, one
can inspect top Facebook groups, top Reddit groups, and top original tweets
(second and third row). Finally, to understand how conversation unfolds, one
can check the retweet and reply networks (bottom row).
After the cleanup, the dataset has 4178 unique URLs. We then use the coun-
try column to select URLs whose column value is “United States.” In the end
our IFCN dataset has 501 URLs.
Tracing Information Cross-platform After we compile two datasets, we
use Information Tracer to collect posts containing those URLs from three plat-
forms, as described in Section 2. For Twitter collection, we set min retweets=10,
min replies=2, max originals=50, and max replies=max retweets=20,000. The
two minimum thresholds filter out low-information tweets, and the three max-
imum thresholds prevent us from burning API quota. Finally, to calculate the
Breakout Scale, we define the breakout threshold to be 100, which means a
URL is considered popular on a platform if its total number of interactions from
that platform is above 100. We experimented different thresholds and found the
resulting trends to be consistent.
3.2 Application 1 : understanding how fake news URLs spread
across platforms
Application 1 demonstrates the core utility of Information Tracer, which is its
ability to quantify information spread over multiple platforms. To facilitate fur-
ther discussion, we categorize each fake news URL into one of four origins: Twit-
ter, Facebook, YouTube and News domains, and consider how URL from each
origin is shared on three platforms – Twitter, Facebook, Reddit. Here, Twitter
and Facebook can be both origin and destination platforms. When we say
the origin of URL A is Twitter, we simply mean A is created on Twitter (i.e.,
A is a tweet). When we say URL B breaks out on Twitter, we mean there is
a high number of tweets that contain URL B, while B can originate from any
platform. Table 2 shows number of URLs from each origin in IFCN and Google
datasets. Specifically, the definition of each origin is:
1. Twitter. URL has a pattern twitter.com/username/tweetid
2. Facebook. URL has a pattern facebook.com/username/type/id, or facebook.com/photo?fbid=id.
type can be posts or videos.
3. Youtube. URL is a YouTube video. For example: youtube.com/watch?v=videoid.
4. News domain. URL is a news article. For example: breitbart.com/article
Given this taxonomy and our multi-dimensional indicators, we investigate
three questions regarding fake news URLs from different origins. We start by an-
alyzing multi-platform patterns (using Breakout Scale and Total Interaction),
then comparing single-platform traffic pattern (using CTM), and finally un-
derstanding contents of fake news from each origin using unsupervised topic
modeling.
Q1: do URLs from different origins have different likelihoods of break-
ing out over multiple platforms? Using the Breakout Scale, we plot the per-
centage of URLs within each origin that spread on 0, 1, 2 and 3 platforms, shown
Table 2: Overview of IFCN and Google datasets, separated by origins of the
URL.
Dataset # URLs # URLs # URLs # URLs Total
Twitter Facebook Youtube News
IFCN FN 65 197 47 192 501
Google FN 241 747 127 312 1427
Table 3: Comparison of median impact indicators of URLs from different origins.
URLs from Facebook are the least likely to spread over multiple platforms.
Dataset avg. tweet per % tweets from top breakout total interactions
user 10% users scale
(I)Twitter 1.05 16 1 994
(I)FB 1.03 14 0 0
(I)Youtube 1.08 18 0 1637
(I)News 1.06 16 1 2080
(G)Twitter 1.06 15 1 544,444
(G)FB N/A N/A 0 0
(G)Youtube 1.04 18 0 272
(G)News 1.07 17 1 251
in Figure 3. We find fake news URLs originating from Facebook (Facebook pages
or images) are the least likely to spread over two or more platforms. Specifically,
more than 90% of URLs from Facebook do not break out on other platforms. In
contrast, 40% of URLs from Twitter, YouTube and News domains break out on
more than one platform, and 20% of URLs from Twitter and YouTube break out
on all three major platforms. This suggests that when fake news is generated on
Facebook, it is more likely to stay within the platform. When a fake news URL
travels across platforms, it is more likely to be a tweet, a YouTube video, or a
news article.
Q2: do URLs from different origins receive different number of inter-
actions and different Twitter traffic? To answer this question, we calculate
the median value of Total Interaction and Coefficient of Traffic Manipulation
(CTM), listed in Table 3. We use median value instead of mean because the
distribution of each indicator is skewed by extremely large values.
The value of total interaction is heavily influenced by the scale and other
aspects of the underlying dataset and API availability. For example, in IFCN
and Google FN datasets, Facebook URLs have a median total interaction of
zero, which indicates that more than half Facebook URLs come from Facebook
groups with low interactions and are not indexed by Crowdtangle API. In ad-
dition, in the Google FN dataset, fake news URLs from Twitter have a median
total interaction of 544,444, a value way larger than total interaction from other
origins. Upon further investigation, we find that most of tweets in the dataset
spread political fake news, and are created by high-profile accounts that receive
unusually high number of interactions, such as @realDonaldTrump (suspended
account of former President Donald Trump, 88 million followers at the time of
suspension) and @seanhannity (TV Host for Fox News, 5.3 million followers as
of February 2021).
For CTM, we do not find any difference among URLs from different origins.
Specifically, median values of average-tweet-per-user range from 1.03 to 1.08, and
median values of percent-tweets-from-top-10%-users range from 14 to 18. The
fact that there is no difference on the aggregated level does not mean no difference
on the individual level. In Section 3.3 we show how to identify individual URL
whose indicators deviate from the norm.
Fig. 3: Percentage of fake news URLs that break out on 0, 1, 2, or 3 platforms,
separated by origin. If the origin is a News domain, YouTube or Twitter, a URL
is more likely to spread over two or more platforms.
Q3: do fake news URLs from different origins cover different topics?
To better understand the substance of fake news, we investigate whether URLs
from different origins cover different topics. To quantify topics, we use non-
negative matrix factorization (NMF), an unsupervised clustering algorithm that
factorizes a document-word matrix into a document-topic matrix and a word-
topic matrix. Using both matrices, we can identify top words within each topic,
and top topics a document belongs to. Previous work has used NMF to discover
meaningful political topics from tweets censored by Turkish government [20].
In both IFCN and Google datasets, there is a “claim” column that summa-
rizes the content of each false URL. The input to the NMF algorithm is thus
a claim-word matrix, where each row is a claim, and each column is a unique
word. The cell value is the tf-idf weight of the word. We lower-case all words,
choose a dictionary size of 5,000 (that is, our matrix has 5,000 columns in max-
imum), and remove all English stopwords. We use Python Scikit-learn package
Fig. 4: Percentage of fake news URLs that belongs to each topic, separated by
origins of the URL. Each color represents a topic. The legend shows keywords
that are most likely to appear within each topic. Topics are discovered using non-
negative matrix factorization. We find that fake news originating from different
platforms cover different topics. For example, in IFCN dataset, topic “5G causes
coronavirus” is more discussed on YouTube than on other platforms, percentage-
wise.
to calculate tf-idf and run NMF [4]. We experiment with different number of
topics, and find that clustering claims of URLs into 6 topics give us meaningful
and interpretable results.
Figure 4 shows proportion of fake news URLs belonging to each topic, and
most frequent words per topic. We find that in the IFCN dataset, topic “5G
causes coronavirus” is a popular YouTube topic (accounting for 14% of all
YouTube URLs), “China bioweapon” is popular within news websites (35% of
news URL), and Twitter users talk more about “President Trump” and his “ad-
ministration.” In the Google FN dataset, we find more discussion about “election
fraud” on Twitter (30% of Twitter URL), and more references to“amala Har-
ris” on YouTube and Facebook. Those differences suggest that fake news topics
are platform-specific. Instead of relying on a fixed list of keywords, social media
platforms can adopt similar methods to discover unknown topics from suspicious
URLs, and use discovered keywords as technical signals to track and stop fake
news spread at an early stage.
3.3 Application 2 : investigating news stories with unusual spread
patterns
Fig. 5: Multi-dimensional visualization of impact indicators. Each marker repre-
sents one URL. Its color reflects the Breakout Scale. Its text reflects origin of
the URL. Its size is in proportion to the total interaction the URL received in
logarithmic scale.
The previous section shows how our framework can compare news spread
patterns of various groups of URLs. Even though aggregated analysis is helpful
to reveal trends or patterns, investigators may also want to examine individual
data points. In this section we show a case study that uses Information Tracer to
understand a URL whose impact indicator deviates from the sample mean. The
URL (denoted as u1 ) we consider is a YouTube video from our IFCN dataset. The
link to the video is https://youtube.com/watch?v=zFN5LUaqxOA. The video
falsely claims that coronavirus is caused by 5G. Though the video has been
removed by YouTube, tweets containing the link are still available. This URL
has an average-tweet-per-user (part of CTM) value of 2, the highest number
among all URLs in IFCN dataset. Figure 5 shows that the URL (inside the red
circle) is on the top-right quadrant, a clear outlier.
To understand why u1 has a high CTM, we navigate to its detail page6 ,
study its retweet network, and find several accounts that repeatedly sent u1 to
targeted users. For example, Figure 6 shows Twitter user @erlhel sharing u1 with
verified accounts, while encouraging users to watch u1 . This spammy behavior
boosts the average-tweet-per-user count. Even though we can not assess whether
account @erlhel is human or bot, its behavior requires more intervention such
as account warning or account suspension.
6
The detail page is available on our web interface:https://informationtracer.com/
?url=youtube.com/watch?v=zFN5LUaqxOA. We encourage investigators to explore
the retweet and reply networks.
Fig. 6: Screenshots of two reply chains. Using Information Tracer, we find account
@erlhel replied under multiple verified accounts, encouraging users to watch
a YouTube video that falsely claims that coronavirus is caused by 5G. This
repetitive tweeting pattern results in a high coefficient of traffic manipulation
(CTM).
3.4 Application 3 : assessing quality of unknown news domains
Another promising use case for Information Tracer is to facilitate human fact-
checking. To assess the utility of our framework, we recruit 30 native English
speakers from surgehq.ai, a platform that provides high-skill workforce. We
ask coders to assess veracity of potentially fake news domains. To obtain such
a list of domains, we adopt and deploy a proactive fake news discovery system,
developed by [8]. The system works in two steps. In step one, it collects live
tweets from Twitter Streaming API based on pre-defined keywords, extracts
domains embedded in tweets, clusters domains together if they are shared by
similar users, and identifies the cluster most related to political news. In step
two, the system assigns a fakeness score to each domain based on a pre-trained
supervised classifier. Our deployed system collected tweets containing keyword
“election” from October 29, 2020 to November 11, 2020. We set a clustering
threshold of 0.6, and selected top 30 unlabeled domains sorted by the fakeness
score. A detailed list of discovered domains is available at https://zhouhanc.
github.io/misinformation-discoverer/. For each domain, we used Informa-
tion Tracer to collect social media posts containing URLs from that domain, and
visualized results on our web interface. For instance, Figure 2 is the screenshot
of social media presence of one discovered domain armyfortrump. com .
We then randomly assigned one domain to one coder, and asked everyone
to assess the factualness of the domain with the help of Information Tracer.
Specifically, we asked coders to look for following signals: is the domain shared
across multiple platforms (e.g., Facebook, Twitter, Reddit)? If so what groups
are sharing the domain on each platform? What hashtags do they use, are they
verified? To teach coders how to navigate through our web interface, we also
shared with them a detailed video instruction7 .
7
The 5-minute video instruction is available on Google Drive: https://drive.
google.com/file/d/1Hqaql5MHlyUKWAwmF7_uKCNyg_ed_nfB/view?usp=sharing
A comprehensive analysis of the accuracy of our deployed fake news discovery
system is beyond the scope of this paper. The result we want to highlight is
people’s perceived utility of Information Tracer. According to Figure 7, when
asked “how helpful is Information Tracer,” 93% coders find it at least somewhat
helpful, and 57% find it very helpful. In addition, we asked coders if they had
any feedback about Information Tracer. One said “I like it a lot except the node
part is really hard to understand”; the other pointed out “It was easier to look
at the page/Twitter page itself, but the recent tweets on Information Tracer gave
a good idea of what the site would be.” We will improve our system based on
those suggestions.
Fig. 7: Perceived utility of Information Tracer from 30 human coders who use the
system to assess factualness of news domains. 93% coders find the tool helpful.
4 Limitation and Next Steps
Data access remains a bottleneck. Despite recent collaboration between
academia and social media platforms, getting access to more accurate metrics
remains a challenge. For example, [14] points out that social media platforms
aggregate two types of metrics – impressions and expressions. Impressions are
publicly available statistics such as number of retweets, replies and likes. Ex-
pressions are more fine-grained measurements such as “who scrolls what tweet
thread for how many seconds.” Impressions can be a better proxy to estimate
the popularity of a post and to derive Breakout Scale. Unfortunately, current
API does not expose impressions data. We hope to engage with platforms and
deepen current collaboration.
Observational data versus experimental data. Even if we can collect all
social media posts, a gap remains where people’s online actions do not necessar-
ily translate to real-world behavioral changes. For example, a story that receives
more interactions may or may not change more people’s behaviors. To measure
behavioral change, controlled experiments are often required. We plan to intro-
duce our framework to the broader political behavior research community. We
also plan to collect alternative data sources such as direct web traffic log or
responses from human subjects to validate our observation.
5 Related Work
Information tracking tools Many open-source tracking systems have been
built over the years. For example, Hoaxy is a system to visualize the spread of
fact-checking claims [16]. FakeNewsTracker [18] is a similar framework to col-
lect, analyze, and visualize tweets related to fake news claims. More recently, [17]
build dashboard to analyze COVID-19 misinformation. [9] provides a more de-
tailed list of open source tools that track misinformation. Limitations of current
frameworks are (1) only focusing on a single platform (usually Twitter) and (2)
not providing sufficient metrics to assess the impact of different news stories.
Our framework aims to overcome those limitations.
Cross-platform misinformation spread Research shows that misinformation
are increasingly spread over multiple platforms. Understanding where misinfor-
mation originates, and where it gets amplified, can help researchers design effec-
tive mitigation strategies [14]. Recently, [21] analyzes the disinformation cam-
paign targeting the White Helmets group using Twitter and Youtube data. [7]
studies how different types of news spread on 4chan and Reddit. [13] collected
URLs from four platforms, Facebook, Twitter, Reddit, and 4chan, quantified
information diffusion, and measured the impact of content moderation. As sug-
gested by [22], previous research in tracing cross-platform news spread lacks a
unified data collection pipeline and well-defined metrics. Our framework aims to
fill this gap.
6 Conclusion
In this paper, we propose and implement Information Tracer, a framework to
track and quantify information spread across multiple platforms. We operational-
ize three metrics – Total Interaction, Breakout Scale and Coefficient of Traffic
Manipulation, and apply our framework on real world datasets. We find that
fake news URLs with different origins have different likelihoods to spread over
multiple platforms, with URLs from Facebook being the least likely to spread
over multiple platforms. Finally, our real-world use cases demonstrate that In-
formation Tracer can help investigators to identify abnormal spread patterns,
facilitate fact-checking, and design better intervention strategies.
References
1. Global Online Content Consumption Doubles in 2020. https://doubleverify.
com/newsroom/global-online-content-consumption-doubles-in-2020-research-shows/
(2020), [Online; accessed February 22, 2021]
2. Download Fact Check Data. https://datacommons.org/factcheck/download
(2021), [Online; accessed April 11, 2021]
3. Google Fact Check Tools API. https://developers.google.com/fact-check/
tools/api (2021), [Online; accessed April 11, 2021]
4. Scikit-learn API Reference. Non-Negative Matrix Factorization (NMF). https:
//scikit-learn.org/stable/modules/generated/sklearn.decomposition.
NMF.html (2021), [Online; accessed April 11, 2021]
5. Berriche, M., Altay, S.: Internet users engage more with phatic posts than
with health misinformation on facebook. Palgrave Communications 6(1), 71
(2020). https://doi.org/10.1057/s41599-020-0452-1, https://doi.org/10.1057/
s41599-020-0452-1
6. Bild, D.R., Liu, Y., Dick, R.P., Mao, Z.M., Wallach, D.S.: Aggregate characteriza-
tion of user behavior in twitter and analysis of the retweet graph. ACM Trans.
Internet Technol. 15(1) (Mar 2015). https://doi.org/10.1145/2700060, https:
//doi.org/10.1145/2700060
7. BURTON, A.G., KOEHORS, D.: Research note: The spread of political misin-
formation on online subcultural platforms (2021), https://doi.org/10.37016/
mr-2020-40
8. Chen, Z., Freire, J.: Proactive discovery of fake news domains from real-time
social media feeds. In: Companion Proceedings of the Web Conference 2020.
p. 584–592. WWW ’20, Association for Computing Machinery, New York,
NY, USA (2020). https://doi.org/10.1145/3366424.3385772, https://doi.org/
10.1145/3366424.3385772
9. Corporation, T.R.: Tools That Fight Disinformation Online . https://www.rand.
org/research/projects/truth-decay/fighting-disinformation/search.html
(2019), [Online; accessed 13-January-2021]
10. Moscadelli, A., Albora, G., Biamonte, M.A., Giorgetti, D., Innocenzio, M.,
Paoli, S., Lorini, C., Bonanni, P., Bonaccorsi, G.: Fake news and covid-
19 in italy: Results of a quantitative observational study. International
journal of environmental research and public health 17(16), 5850 (08
2020). https://doi.org/10.3390/ijerph17165850, https://pubmed.ncbi.nlm.nih.
gov/32806772
11. Nimmo, B.: Measuring traffic manipulation on twitter (2021), https:
//comprop.oii.ox.ac.uk/wp-content/uploads/sites/93/2019/01/
Manipulating-Twitter-Traffic.pdf, [Online; accessed January 5, 2021]
12. Nimmo, B.: The breakout scale: Measuring the impact of influence op-
erations (2021), https://www.brookings.edu/wp-content/uploads/2020/09/
Nimmo_influence_operations_PDF.pdf, [Online; accessed January 5, 2021]
13. PAPAKYRIAKOPOULOS, O., SERRANO, J.C.M., HEGELICH, S.: The spread
of COVID-19 conspiracy theories on social media and the effect of content moder-
ation (2021), https://doi.org/10.37016/mr-2020-034
14. Pasquetto, I.V., Swire-Thompson, B., Amazeen, M.A., et al.: Tackling misin-
formation: What researchers could do with social media data (2021), https:
//doi.org/10.37016/mr-2020-49
15. Pulido, C.M., Ruiz-Eugenio, L., Redondo-Sama, G., Villarejo-Carballido, B.: A
new application of social impact in social media for overcoming fake news in
health. International journal of environmental research and public health 17(7),
2430 (04 2020). https://doi.org/10.3390/ijerph17072430, https://pubmed.ncbi.
nlm.nih.gov/32260048
16. Shao, C., Ciampaglia, G.L., Flammini, A., Menczer, F.: Hoaxy. Proceedings of
the 25th International Conference Companion on World Wide Web - WWW ’16
Companion (2016). https://doi.org/10.1145/2872518.2890098, http://dx.doi.
org/10.1145/2872518.2890098
17. Sharma, K., Seo, S., Meng, C., Rambhatla, S., Liu, Y.: Covid-19 on social media:
Analyzing misinformation in twitter conversations (2020)
18. Shu, K., Mahudeswaran, D., Liu, H.: Fakenewstracker: a tool for fake news col-
lection, detection, and visualization. Computational and Mathematical Organi-
zation Theory 25(1), 60–71 (2019). https://doi.org/10.1007/s10588-018-09280-3,
https://doi.org/10.1007/s10588-018-09280-3
19. Statista: Most popular mobile social networking apps in the united states
as of september 2019 (2020), https://www.statista.com/statistics/248074/
most-popular-us-social-networking-apps-ranked-by-audience/, [Online; ac-
cessed January 8, 2021]
20. Tanash, R.S., Chen, Z., Thakur, T., Wallach, D.S., Subramanian, D.:
Known unknowns: An analysis of twitter censorship in turkey. In: Pro-
ceedings of the 14th ACM Workshop on Privacy in the Electronic Soci-
ety. p. 11–20. WPES ’15, Association for Computing Machinery, New York,
NY, USA (2015). https://doi.org/10.1145/2808138.2808147, https://doi.org/
10.1145/2808138.2808147
21. Wilson, T., Starbird, K.: Cross-platform disinformation campaigns: lessons learned
and next steps (2021), https://doi.org/10.37016/mr-2020-002
22. Yang, K.C., Pierri, F., Hui, P.M., Axelrod, D., Torres-Lugo, C., Bryden, J.,
Menczer, F.: The covid-19 infodemic: Twitter versus facebook (2020)
23. Zarocostas, J.: How to fight an infodemic. The Lancet 395(10225), 676
(2021/02/19 2020). https://doi.org/10.1016/S0140-6736(20)30461-X, https://
doi.org/10.1016/S0140-6736(20)30461-X